How to conduct out-of-sample prediction

[MichaelHopwood]

Is there an established way to utilize GraphSage's inductive properties to make an inference on a new graph (e.g. the same graph in training but with more nodes added)?

The examples I see in the examples/ directory utilize an index to cut up a single graph for train & test sets.

As a note - I'd want to apply this for either a node classification problem or an unsupervised node embedding problem.

[rusty1s]

Yes, you should be able to re-use the same node but on a modified graph. For this, you can simply create a new loader that now operates on your new data. Let me know if this helps you get started!

[rusty1s]

Yes, this looks good. Please note that you also need to add reverse edges as well:

[24, 2708], [245, 2708], [622, 2708]

[MichaelHopwood]

@rusty1s Thanks for your response. Is there anything inhibiting inference on a completely new data?

For instance, I trained a model on my data and then did an inference on Cora. Other than mandating that the node features maintains the same feature dimension, it seemed to work fine.

# model = trained on my dataset, using inductive & NeighborLoader

# Start inference
dataset_cora = get_planetoid_dataset('cora')
dataset = [Data(x=dataset_cora[0].x[:, :41].clone(),
                y=dataset_cora[0].y,
                edge_index=dataset_cora[0].edge_index)]
preds = model(dataset[0])
# works without error

Does doing this raise any red flags?
Is this how transfer learning is conducted in GNNs?

[rusty1s]

I am not sure if this makes a lot of sense. The domain and features in your data need to match with the ones the model was trained on. Otherwise, this just corresponds to an untrained model.

[MichaelHopwood]

Yeah sorry I just meant this as a proof of concept. Certainly it doesn't make sense here but if I have (for instance) a graph which updates every few days and shares the same feature set, then I suppose I would use this method to perform inductive inference on the new graph & nodes.

[rusty1s]

Yes, that is correct.